JP2002230552A - Pattern recognition process device and method and pattern recognition program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a pattern recognition process device and method and a pattern recognition program capable of causing an overall atmosphere to reflect on the result of recognition while making the burden of a process relatively light. SOLUTION: An averaging process part 14 divides an original image into a plurality of regions and averages a gradation value for each region. A two-dimensional FFT process part 16 performs a two-dimensional FFT process on the original image and another averaging process part 18 divides the result into a plurality of regions and performs averaging for each of the regions. A plurality of average values obtained by the two averaging process parts 14 and 18 are input to a neural network process part 20, and of learning results the one closest to the real pattern is extracted as the result of recognition.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a pattern recognition processing apparatus, a method, and a pattern recognition program for photographing various subjects such as a human body, a landscape, and small objects using a camera and performing pattern recognition on the subjects.

[0002]

2. Description of the Related Art Conventionally, FA (Factory Automation)
In such a field, a pattern recognition technique for extracting a partial image corresponding to a specific target object and recognizing the content based on an image captured by a camera is used. For example, when an operation of sorting a large variety of small products is performed unattended, a pattern recognition technology is used to identify the types of products transferred from a previous process.

[0003]

In the conventional pattern recognition technique, an image to be recognized is cut out of a picked-up image, and pattern comparison is performed on the cut out image. Therefore, when performing pattern recognition, it is also necessary to always perform image cutout processing,
There was a problem that the processing load was heavy. In addition, since only the cut-out portion is subjected to pattern recognition, there is a problem that the atmosphere of the entire imaging range cannot be reflected on pattern recognition.

[0004] The present invention has been made in view of the above points, and its object is to relatively lightly burden the processing.
An object of the present invention is to provide a pattern recognition processing device, a method, and a pattern recognition program that can reflect the overall atmosphere in a recognition result.

[0005]

In order to solve the above-mentioned problems, a pattern recognition processing apparatus according to the present invention comprises first and second averaging processing means, frequency analysis means, neural network processing means, and output means. Have. The first averaging unit divides the first image of the object to be recognized into a plurality of regions, and averages the gradation values for each of the divided regions. The frequency analysis means performs a frequency analysis using the values of each pixel constituting the first image. The second averaging unit divides the two-dimensional distribution of the analysis result obtained by the frequency analysis unit into a plurality of regions, and averages the analysis result for each divided region. The neural network processing means comprises:
And the respective average values obtained by the second averaging processing means are input in correspondence with a plurality of cells included in the input layer.
The output means obtains a recognition result by extracting an output value of a plurality of cells included in an output layer of the neural network processing means which is close to a target. The first image obtained by capturing the object to be recognized and the two-dimensional distribution obtained as a result of frequency analysis of the image are divided into a plurality of regions, and the result of calculating the average value for each divided region is stored in a neural network. By inputting, pattern recognition is performed, and a process of cutting out a specific recognition target image becomes unnecessary, so that the processing load can be reduced. In addition, since pattern recognition is performed using the entire imaging range, the overall atmosphere can be reflected in the recognition result. Furthermore, since pattern recognition is performed using not only a captured image but also a result of frequency analysis of the image, recognition accuracy can be improved.

It is desirable that the above-mentioned target is a maximum value or a minimum value. It is easier to check whether the value is the maximum value or the minimum value than to check whether the value is close to a specific value (target value). Speeding up becomes possible.

[0007] Further, there are provided an image pickup means for picking up an image of the object to be recognized, and a preprocessing means for performing a predetermined process on an original image obtained by the image pickup by the image pickup means to generate a first image. It is desirable. Since preprocessing is performed on the first image actually captured, image data (for example, grayscale image data) suitable for subsequent processing can be obtained, and processing can be simplified. Become.

It is desirable that the frequency analysis performed by the above-described frequency analysis means be a two-dimensional fast Fourier transform process. By the two-dimensional fast Fourier transform processing, an analysis result having a planar spread can be easily obtained, so that the characteristics of the recognition target can be grasped more accurately, and the recognition accuracy can be improved. .

Further, the pattern recognition processing method of the present invention comprises:
A first step of dividing the first image of the object to be recognized into a plurality of regions, averaging the gradation values for each of the divided regions, and using a value of each pixel constituting the first image. A second step of performing a frequency analysis by performing the above, a third step of dividing the two-dimensional distribution of the analysis result obtained in the second step into a plurality of regions, and averaging the analysis result for each divided region; A fourth step of inputting each of the average values obtained in the first and third steps in correspondence with a plurality of cells included in the input layer and outputting output values of the plurality of cells included in the output layer of the neural network; And a fifth step of obtaining a recognition result by extracting a value close to the target from among the plurality of output values output in the fourth step. Further, the pattern recognition program of the present invention comprises:
In order to recognize the contents of the recognition target, the computer
A first step of dividing the first image of the object to be recognized into a plurality of regions and averaging the gradation values for each of the divided regions, using a value of each pixel constituting the first image; A second step of performing a frequency analysis, a third step of dividing the two-dimensional distribution of the analysis result obtained in the second step into a plurality of regions, and averaging the analysis result for each of the divided regions; A fourth step of inputting each of the average values obtained in the third step in association with a plurality of cells included in the input layer and outputting output values of the plurality of cells included in the output layer of the neural network; The fifth step is to obtain a recognition result by extracting a plurality of output values output in the fourth step that are close to the target.

By implementing the pattern recognition processing method of the present invention or executing the pattern recognition program of the present invention on a computer, pattern recognition can be performed without cutting out a specific image to be recognized. Burden can be reduced. Further, by performing pattern recognition using the entire imaging range, the overall atmosphere can be reflected in the recognition result. Furthermore, since pattern recognition is performed using not only a captured image but also a result of frequency analysis of the image, recognition accuracy can be improved.

[0011]

DESCRIPTION OF THE PREFERRED EMBODIMENTS A pattern recognition processing apparatus according to an embodiment of the present invention will be described below in detail with reference to the drawings. FIG. 1 is a diagram illustrating a configuration of a pattern recognition processing device according to the present embodiment. The pattern recognition processing device according to the present embodiment shown in FIG. 1 includes a camera 10, a pattern recognition processing unit 40, and an information storage medium 50.

The camera 10 captures an image of a recognition target 100 placed on the stage 30. For example, a CCD (Charge Coupled Device) is used as an image sensor. The pattern recognition processing unit 40 specifies the content of the recognition target 100 captured by the camera 10 by performing a predetermined pattern recognition process. For this purpose, the pattern recognition processing unit 40 includes the preprocessing unit 12, the averaging processing units 14 and 18, the two-dimensional FFT processing unit 16, the neural network processing unit 2
0, a maximum value extracting unit 22, and a learning input unit 24.

The pre-processing unit 12 converts image data consisting of three color data output from the camera 10 into image data consisting of grayscale gradation data, and converts the image data after the conversion process. To perform predetermined pre-processing. As the pre-processing, processing for correcting unevenness in brightness that occurs when the stage 30 is imaged by the camera 10 can be considered, but this processing may be omitted if the degree of unevenness is small. .

The averaging processing unit 14 receives grayscale tone values corresponding to each pixel included in the imaging range from the preprocessing unit 12, and divides the imaging range into a plurality of regions. The average of the gradation values for each divided area is calculated. Here, a so-called mosaic process is performed on the grayscale original image.

A two-dimensional FFT (fast Fourier transform) processing unit 16 performs a two-dimensional FFT process on a grayscale two-dimensional image output from the preprocessing unit 12. Also,
The averaging unit 18 divides the two-dimensional distribution of the analysis result by the two-dimensional FFT unit 16 into a plurality of regions, and calculates an average for each of the divided regions. When the two-dimensional distribution of the analysis result is represented in the form of a two-dimensional image, the processing by the averaging processing unit 18 also corresponds to a so-called mosaic processing, similar to the processing by the averaging processing unit 14 described above.

The neural network processing unit 20
Output values are output from a plurality of cells included in the output layer of the neural network by making a plurality of average values obtained by the two averaging processing units 14 and 18 correspond to a plurality of cells of the input layer.
The maximum value extraction unit 22 is a neural network processing unit 2
A value having the maximum value is extracted from a plurality of output values of 0, and a cell at this time is specified to derive a recognition result. As a premise to derive this recognition result,
It is necessary to make the neural network learn the object to be recognized. Various methods are known as algorithms for learning a neural network. For example, it is assumed that learning is performed using a BP method (back propagation method) or the like.

The learning input unit 24 is used to designate a specific cell in the output layer in order to make a correct recognition result remembered when learning the neural network.
The information storage medium 50 stores a program necessary for performing the pattern recognition processing performed by the pattern recognition processing unit 40. This information storage medium 50
It is configured using an optical disk-type storage medium such as a CD (compact disk), a semiconductor memory such as a ROM or a RAM, or a hard disk device.

For example, as the information storage medium 50, a ROM or RA storing a series of programs for pattern recognition processing is stored.
When M is considered, the preprocessing unit 12, the averaging units 14, 18 included in the pattern recognition processing unit 40, and the two-dimensional FFT
The processing unit 16, the neural network processing unit 20, the maximum value extracting unit 22, and the learning input unit 24
It can be realized by executing by a PU.

The above-mentioned camera 10 serves as an image pickup means, the preprocessing section 12 serves as a preprocessing means, the averaging processing section 14 serves as a first averaging processing means, the two-dimensional FFT processing section 16 serves as a frequency analysis means, and The processing unit 18 corresponds to the second averaging processing unit, the neural network processing unit 20 corresponds to the neural network processing unit, and the maximum value extracting unit 22 corresponds to the output unit.

The pattern recognition processing apparatus of the present embodiment has such a configuration, and the processing procedure will be described next. FIG. 2 shows the recognition target 1 placed on the stage 30.
9 is a flowchart showing a series of pattern recognition procedures from capturing an image of 00 by the camera 10 to recognizing the contents. Hereinafter, a series of pattern recognition operations will be described with reference to FIG.

First, the stage 30 is controlled by the camera 10.
An image of the recognition target 100 arranged above is taken (step 100). The camera 100 outputs color data (RGB data) of each pixel of the captured image. Preprocessing unit 12
Calculates grayscale gradation data based on the color data output from the camera 10 (step 10).
1).

Next, the averaging processing unit 14 performs a predetermined averaging process, which is generally called a mosaic process, on the original image converted into the gray scale (step 102).
In parallel with or before or after this averaging process,
The two-dimensional FFT processing unit 16 performs two-dimensional FFT processing on the original image converted to gray scale (step 1).
03). After that, the averaging processing unit 18 outputs the two-dimensional FF
A predetermined averaging process is performed on the result of the T process (step 104).

Next, the neural network processing unit 20
Performs a predetermined process using a neural network by associating a plurality of average values obtained in each of steps 102 and 104 with each cell of the input layer (step 105). At this time, the maximum value extracting unit 22
The cell having the maximum value is extracted from the plurality of cells included in the output layer of the neural network (step 10).
6) Output the recognition result corresponding to this cell (step 107).

FIG. 3 is a diagram schematically showing the flow of the pattern recognition processing described above. 3, an averaging process indicated by B (Step 102 in FIG. 2) and a two-dimensional FFT process indicated by C are performed based on the grayscale original image indicated by A (103 in FIG. 2). . Then, based on the result of the two-dimensional FFT processing, an averaging processing indicated by D (Step 104 in FIG. 2) is performed.

In this way, a result of averaging the original image itself and a result of averaging the two-dimensional FFT processed image of the original image are obtained.
The result is input to the neural network indicated by E, and one recognition result with the highest probability of being finally obtained is finally obtained.

In FIG. 3, for example, a neural network of 96 cells in the input layer, 30 cells in the intermediate layer, and 10 cells in the output layer is used, but the number of layers and the number of cells in each layer can be arbitrarily changed. As described above, the pattern recognition processing apparatus according to the present embodiment divides each of the grayscale original image and the two-dimensional distribution resulting from the two-dimensional FFT processing of the original image into a plurality of regions, and By inputting the result of calculating the average value into the neural network, pattern recognition is performed, and the processing of cutting out a specific recognition target image becomes unnecessary, so that the processing load can be reduced. Further, since pattern recognition is performed using the entire imaging range, the atmosphere around the recognition target can be reflected in the recognition result. Furthermore, not only the result of averaging the original image but also the two-dimensional FF
By using the result of the averaging process after the T process, the accuracy of pattern recognition can be improved.

The present invention is not limited to the above embodiment, and various modifications can be made within the scope of the present invention. For example, in FIG. 1, the case where the recognition target object 100 is arranged on the stage 30 is considered. However, as shown by an original image A in FIG. 3, pattern recognition is performed based on an image obtained by capturing an arbitrary space. May be performed. Further, it is not always necessary to perform pattern recognition using the entire imaging range, and pattern recognition may be performed by extracting a part of the image of the imaging range.

In the above-described embodiment, the maximum value is extracted from the plurality of output values of the neural network processing unit 20 to obtain the recognition result. Alternatively, a recognition result may be obtained by extracting a result similar to the above. However, the process of detecting the maximum value or the minimum value is lighter in load than the process of detecting a value close to the intermediate target value, and is therefore preferable in reducing the load or speeding up the process.

[0029]

As described above, according to the present invention, each of the first image of the object to be recognized and the two-dimensional distribution resulting from the frequency analysis of this image is divided into a plurality of regions. By inputting the result of calculating the average value for each divided region to the neural network, pattern recognition is performed, and the process of cutting out a specific image to be recognized becomes unnecessary, so that the processing load can be reduced. it can. Also,
Since pattern recognition is performed using the entire imaging range,
The overall atmosphere can be reflected in the recognition result.
Furthermore, since pattern recognition is performed using not only a captured image but also a result of frequency analysis of the image, recognition accuracy can be improved.

[Brief description of the drawings]

FIG. 1 is a diagram illustrating a configuration of a pattern recognition processing device according to an embodiment.

FIG. 2 is a flowchart showing a series of pattern recognition procedures from capturing an image of a recognition target placed on a stage using a camera to recognizing the contents thereof.

FIG. 3 is a diagram schematically showing a flow of a pattern recognition process.

[Explanation of symbols]

 DESCRIPTION OF SYMBOLS 10 Camera 12 Preprocessing part 14, 18 Averaging part 16 Two-dimensional FFT processing part 20 Neural network processing part 22 Maximum value extraction part 24 Learning input part 40 Pattern recognition processing part 50 Information storage medium

Claims (6)

[Claims] A first averaging unit that divides a first image of the object to be recognized into a plurality of regions, and averages a gradation value for each divided region; and the first image. Frequency analysis means for performing frequency analysis using the value of each pixel constituting the, the two-dimensional distribution of the analysis result obtained by the frequency analysis means is divided into a plurality of regions, and the analysis results are averaged for each divided region A second averaging processing means for converting the average value obtained by the first and second averaging processing means into a plurality of cells included in the input layer, and a neural network processing means for inputting the average values. A pattern recognition processing apparatus comprising: an output unit that obtains a recognition result by extracting a value close to a target from output values of a plurality of cells included in an output layer of the neural network processing unit. 2. The pattern recognition processing device according to claim 1, wherein the target is a maximum value or a minimum value. 3. The first image according to claim 1, wherein the first image is generated by performing predetermined processing on an original image obtained by imaging by the imaging unit. A pattern recognition processing device, comprising: 4. The pattern recognition processing device according to claim 1, wherein the frequency analysis performed by the frequency analysis unit is a two-dimensional fast Fourier transform process. 5. A first step of dividing a first image in which a recognition target object is imaged into a plurality of regions and averaging a gradation value for each of the divided regions; and forming the first image. A second step of performing frequency analysis using the value of each pixel; and dividing the two-dimensional distribution of the analysis result obtained in the second step into a plurality of regions, and averaging the analysis results for each of the divided regions. And inputting the average values obtained in the first and third steps in association with a plurality of cells included in an input layer, and a plurality of cells included in an output layer of the neural network. And a fifth step of obtaining a recognition result by extracting a plurality of output values near the target from among the plurality of output values output in the fourth step. Pattern recognition processing characterized by Law. 6. A computer divides a first image of the object to be recognized into a plurality of regions and averages a gradation value for each of the divided regions in order to recognize the contents of the object to be recognized. A first step, a second step of performing frequency analysis using values of each pixel constituting the first image, and dividing a two-dimensional distribution of an analysis result obtained in the second step into a plurality of regions A third step of averaging the analysis results for each of the divided areas; and inputting the respective average values obtained in the first and third steps in association with a plurality of cells included in the input layer. A fourth step of outputting output values of a plurality of cells included in the output layer of the neural network; and extracting a recognition result by extracting a plurality of output values close to a target among the plurality of output values output in the fourth step. The fifth step to get, Pattern recognition program for executing